Datasets: File Name: AlexStrofs_DataExplore.ipynb Neural Network.ipynb
Abstract
This project explores advanced data analysis techniques to enhance healthcare outcomes and understand disease, focusing on Alzheimer’s disease detection through MRI data and cognitive measures. By leveraging clustering, polynomial regression, and a highly accurate neural network model, this research identifies critical cognitive health markers and predictive models for early Alzheimer's diagnosis. The findings underscore the efficacy of MRI biomarkers in predictive healthcare, with a neural network yielding the highest accuracy in cognitive decline predictions. These results suggest that machine learning can support clinicians in early intervention, potentially improving patient outcomes.
Table of Contents
Introduction
Literature Review
Methodology
Neural Network Component
Results
Discussion
Limitations
Key Takeaways
Further Research
Conclusion
Bibliography
Introduction
The potential for advanced data analysis techniques in healthcare is vast. Effective methods enable more accurate diagnoses, improving patient outcomes and optimizing resource management. Machine learning’s predictive power can also forecast critical events, allowing for proactive intervention. This study examines data analysis techniques and their applications in predicting and understanding Alzheimer’s disease.
Literature Review
Predictive Analytics:
This study examines the implementation of predictive analytics in healthcare, focusing on models integrated with electronic health records (EHRs). Liu et al. introduce a framework that addresses the clinical value of predictive models, emphasizing the prediction-action dyad through metrics such as Number Needed to Screen (NNS), Number Needed to Treat (NNT), and a novel metric, Number Needed to Benefit (NNB). They argue that while the deployment of predictive models is on the rise, assessing their real-world clinical benefits poses challenges. The authors recommend a shared framework for the effective and safe use of these models in practice, aiming to optimize patient outcomes.
Key statistical concepts critical for evaluating predictive models include the Number Needed to Screen (NNS), Number Needed to Treat (NNT), and Number Needed to Benefit (NNB). These metrics provide essential insights into the effectiveness of predictive analytics in healthcare. Additionally, performance assessment of these models often relies on the Positive Predictive Value (PPV) and the Area Under the Receiver Operating Characteristic Curve (AUC-ROC), which help determine the accuracy and reliability of predictions. However, the paper raises several points of confusion, notably the lack of clarity regarding the detailed calculation of NNB from NNS and NNT, the derivation of confidence intervals for NNB, and the need for standardized evaluations of predictive models across different healthcare settings(Liu et al. 2019).
Zhang discusses the transformative impact of big data on clinical medicine, highlighting predictive analytics' role across the disease continuum. The author outlines the three critical steps in big data analytics: formulating clinical questions, designing studies, and conducting statistical analysis. Challenges related to high-dimensional data and model interpretability are emphasized, along with the necessity for practical tools in healthcare predictive modeling.
Statistical methods and concepts play a pivotal role in the analysis and interpretation of predictive models in healthcare. Generalized Linear Models (GLMs) and neural networks are frequently discussed in this context, each offering distinct advantages for modeling complex relationships within data. While GLMs provide interpretable results and are suitable for various types of outcome variables, neural networks excel in capturing intricate patterns but often lack transparency, which can pose challenges in clinical settings. Understanding the strengths and limitations of these approaches is essential for effectively utilizing them in predictive analytics.
Validation techniques are crucial in ensuring the robustness of predictive models, with a strong emphasis on external validation to mitigate the risk of overfitting. Overfitting occurs when a model is excessively complex, capturing noise rather than the underlying data trends, leading to poor generalization to new datasets. By prioritizing external validation, researchers can confirm the applicability of their models across different populations and settings, thereby enhancing their credibility and usefulness in clinical practice.
However, several points of confusion arise in the literature. Notably, the classification of neural networks as non-parametric is incorrect, as they are inherently parametric due to their reliance on a fixed number of weights and parameters. Additionally, the paper fails to clearly distinguish between tools designed for model evaluation and those intended for clinical application. This lack of clarity can lead to misunderstandings regarding the appropriate use of these statistical methods in real-world healthcare scenarios. Addressing these confusions is vital for advancing the field and ensuring that predictive analytics are implemented effectively and safely (Zhang 2020).
Visualization Techniques:
This narrative review investigates recent advancements in visualization techniques within healthcare. It emphasizes interactive visualization's role in enhancing data comprehension and the application of AI and machine learning in visualizing health metrics. The authors provide a descriptive analysis of studies from 2018 to 2021, highlighting the importance of visualization in conveying complex statistical data.
Statistical methods and concepts related to data visualization are essential for enhancing user comprehension, particularly through interactive visualization techniques that allow users to manipulate graphical representations of data. These tools significantly improve the interpretability of complex datasets by facilitating a more intuitive understanding of the underlying information. Various statistical software programs provide interactive capabilities that empower users to engage with their data actively, promoting deeper insights and more informed decision-making. However, several points of confusion persist in the literature, particularly regarding the specifics of the statistical methods employed by AI and machine learning tools. There is a notable lack of detail on how the accuracy of interactive visualizations is validated, raising questions about the reliability of these representations in accurately conveying the data's true characteristics. Addressing these gaps is crucial for ensuring the effectiveness of interactive visualization techniques in data analysis (Abudiyab and Alanazi 2022).
Machine Learning Applications:
Machine learning applications have shown significant promise in the areas of disease prediction and progression, particularly through the use of supervised machine learning algorithms. A comprehensive review of 48 articles reveals that Support Vector Machine (SVM) and Naïve Bayes algorithms dominate this field, while Random Forest (RF) emerges as the most accurate predictor of diseases. This analysis aids researchers in selecting the most appropriate algorithm based on comparative performance metrics, emphasizing the importance of accuracy measurement—the ratio of correctly predicted instances to total instances. Additionally, ensemble methods like RF, which combine multiple decision trees, enhance prediction accuracy. However, several areas of uncertainty remain, such as the specifics of accuracy calculations across studies, the statistical significance of findings, and how class imbalances are addressed(Ahsan et al. 2022).
In the context of Alzheimer’s disease (AD), another study employs interpretable machine learning to identify immune microenvironment subtypes among patients. By utilizing various algorithms, including XGBoost, the researchers aim to predict AD outcomes and identify key genes linked to the disease. They incorporate techniques such as Single-sample Gene Set Enrichment Analysis (ssGSEA) and LASSO regression to analyze immune states and improve model interpretability, with performance metrics like Area Under the Curve (AUC) and Precision-Recall (P-R) values used to evaluate model effectiveness. Nevertheless, uncertainties persist regarding the clarity of model training and validation processes, the statistical significance of the results, and the management of class imbalances (Lai et al. 2022).
Another study focuses on predicting 30-day readmissions following ischemic stroke using electronic health record (EHR) data. Through a comprehensive analysis involving multiple machine learning algorithms, key predictors for readmissions are identified, highlighting the critical role of structured EHR data in enhancing patient care quality. Key statistical methods include feature selection and adaptive sampling, which are essential for improving model performance. The evaluation of these models employs various metrics, including AUC, sensitivity, specificity, and positive predictive value (PPV). However, certain unclear concepts, such as the implementation specifics of ROSE-sampling, hyperparameter tuning, and the details of confusion matrices, warrant further clarification to enhance understanding and application in clinical settings(Chen et al. 2022).
The literature review highlighted the growing importance of predictive analytics and machine learning in healthcare, particularly in understanding disease progression and improving diagnostic accuracy. Key papers identified predictive models that leverage electronic health records (EHRs) and MRI data to improve patient outcomes. These studies emphasized the need for effective methodologies that not only classify conditions but also provide insights into the underlying mechanisms of diseases.
Motivated by these insights, I conducted a thorough analysis of datasets derived from the Open Access Series of Imaging Studies (OASIS), focusing specifically on MRI-derived features and their relationship to Alzheimer's disease. This journey led to the application of several machine learning models tailored to predict Alzheimer's disease effectively, allowing for a deeper understanding of the factors influencing cognitive decline.
Methodology
The research uses MRI data from the Open Access Series of Imaging Studies (OASIS) to identify biomarkers linked to Alzheimer’s. Models including K-means clustering, polynomial regression, and a neural network were developed and evaluated on their predictive accuracy.
Neural Network Component
Building a neural network to predict Mini-Mental State Examination (MMSE) scores was a strategic choice driven by both the importance of MMSE as a cognitive assessment tool and the complex relationships observed between MMSE and other brain health indicators in the dataset. The MMSE is widely regarded as a gold-standard measure for evaluating cognitive function, particularly in assessing the severity of cognitive impairment and tracking changes over time. As such, it is invaluable in the early detection of conditions like Alzheimer’s disease, where timely intervention can make a significant difference in patient outcomes.
Through clustering and regression analyses, distinct relationships emerged between MMSE scores and various MRI-derived features, including Atlas Scaling Factor (ASF), normalized Whole Brain Volume (nWBV), estimated Total Intracranial Volume (eTIV), and Clinical Dementia Rating (CDR). For instance, K-means clustering with ASF, nWBV, and eTIV revealed different cognitive profiles, suggesting that these MRI biomarkers relate to cognitive health in complex ways. Individuals with higher brain volume (e.g., higher eTIV and nWBV) tended to show better cognitive function, as measured by MMSE. Similarly, patterns in ASF indicated its potential role as a secondary indicator of brain structure changes, possibly aligning with cognitive states in aging individuals.
However, while these clustering analyses offered insights, they also highlighted the limitations of simpler models in capturing the nuanced, nonlinear relationships between these features and MMSE scores. Polynomial regression models, for example, tended to overfit at higher degrees, capturing noise rather than meaningful patterns in the data. These limitations underscored the need for a more sophisticated approach capable of modeling complex, non-linear relationships across multiple variables.
A neural network was therefore selected as the ideal predictive tool for MMSE due to its ability to learn intricate patterns in the data that might be missed by linear or polynomial models. Neural networks are particularly suited to capturing hidden interactions and non-linear relationships between features, making them valuable for predicting outcomes like MMSE, where cognitive decline may not be linearly or independently related to a single feature. By leveraging input features such as ASF, nWBV, and eTIV, the neural network could potentially identify subtle patterns in how structural brain metrics and clinical assessments collectively influence cognitive function.
Furthermore, the predictive power of MMSE scores, as demonstrated by the clustering results, indicates that they could act as an effective proxy for overall cognitive health. Accurate predictions of MMSE could therefore enable earlier detection of cognitive decline, which is crucial in conditions like Alzheimer’s disease. MMSE prediction aligns with the broader goals of developing tools for early diagnosis, targeted intervention, and personalized treatment plans for individuals at risk of cognitive impairment.
Results
Clustering Analysis
The K-means clustering analysis provided key insights into cognitive health profiles relevant to Alzheimer’s disease. Using the elbow method to determine the optimal number of clusters, K=2 emerged as the best choice, marking a significant decrease in inertia and capturing the primary structure of the dataset. Beyond K=3, additional clusters yielded only marginal improvements, suggesting that dividing the data into two clusters effectively represented its main patterns.
For the clustering analysis focused on Mini-Mental State Examination (MMSE) scores and Atlas Scaling Factor (ASF), two distinct clusters emerged. Cluster 1 primarily included individuals with higher MMSE scores, indicating better cognitive function, regardless of ASF levels. In contrast, Cluster 0 comprised individuals with a broader range of ASF values and generally lower MMSE scores, suggesting diverse or potentially declining cognitive states. This relationship between ASF and MMSE suggests that ASF could serve as an additional metric for understanding variations in cognitive health, especially in populations at risk for Alzheimer’s.
Further, examining the clustering based on MMSE and normalized Whole Brain Volume (nWBV) suggested a positive relationship between brain volume and cognitive function. Cluster 1 contained individuals with higher MMSE scores across a range of nWBV values, likely representing individuals with relatively preserved cognitive function. Cluster 0, however, exhibited greater variability in both MMSE scores and brain volume, indicating a broader spectrum of cognitive states, including potential cognitive decline. These findings suggest that higher nWBV values may align with better cognitive function.
The clustering results have important implications. First, cognitive health segmentation allows for categorizing individuals based on brain measurements and cognitive scores, which could help in identifying patient profiles that might benefit from customized treatment approaches. Additionally, the observed relationships between MMSE, ASF, and nWBV imply that these features could be valuable in early cognitive health assessments, supporting their use in Alzheimer’s screening and diagnosis.
The scatter plot below represents K-means clustering applied to the variables Age and eTIV (Estimated Total Intracranial Volume), with K=2 clusters identified. The clusters are distinguished by different markers: Cluster 0 (teal circles) and Cluster 1 (orange crosses).
Cluster 0 is largely characterized by individuals with lower eTIV values, primarily below 1600, and spans a wide range of ages from approximately 40 to 90 years. This indicates that within Cluster 0, individuals of varying ages tend to have smaller brain volumes. Cluster 1, in contrast, consists mostly of individuals with higher eTIV values, typically above 1600, and is more concentrated within the age range of 70 to 90 years. This pattern suggests that larger brain volumes may be associated with the older age group within this dataset.
The clustering pattern appears to be more influenced by eTIV than by age. While both clusters include a wide age range, the eTIV values provide a clearer distinction between the two groups. This finding implies that eTIV might be a significant differentiating factor within this population, with age playing a secondary role in cluster formation.
The distribution of clusters may have implications for understanding how brain volume (eTIV) and age interact, possibly reflecting differences in brain health or cognitive characteristics among these groups. To gain further insights, additional analysis could be conducted to examine how these clusters correlate with cognitive measures, such as MMSE scores, or other health indicators. This analysis may help identify whether larger brain volumes in older individuals correlate with cognitive preservation or other neurological factors.
The K-means clustering analysis of Age versus normalized Whole Brain Volume (nWBV) with K=2 clusters provides insights into how brain volume changes across age groups in the context of cognitive health. In the scatter plot, two clusters are observed: Cluster 0 (teal circles) and Cluster 1 (orange crosses). These clusters demonstrate a clear trend in the relationship between age and nWBV.
Cluster 1 predominantly comprises younger individuals (ages roughly between 40 and 70) who tend to have higher nWBV values, generally above 0.775. This clustering indicates that younger individuals in this sample are more likely to have larger brain volumes, which aligns with the understanding that brain volume tends to decrease with age due to natural atrophy. The concentration of Cluster 1 among younger participants with higher nWBV values suggests a potential group with preserved cognitive health or less brain atrophy, likely associated with better overall brain function.
Cluster 0, on the other hand, includes primarily older individuals (ages roughly between 70 and 90) with generally lower nWBV values, mostly below 0.775. The presence of older individuals in this cluster with lower nWBV aligns with the expected pattern of age-related brain atrophy, which is commonly observed in aging populations and is often associated with cognitive decline. The data in Cluster 0 suggest a group potentially at higher risk for cognitive impairment, as lower nWBV values can indicate greater brain volume loss.
The clear separation between these clusters underscores the relationship between age and nWBV: as individuals age, there is a tendency for nWBV to decrease. This finding aligns with clinical observations of brain volume reduction with aging and its possible association with cognitive decline. The clustering pattern indicates that nWBV could be a valuable metric in assessing brain health across age groups and in identifying individuals at higher risk for cognitive issues.
This clustering pattern suggests that nWBV, in conjunction with age, may serve as an effective biomarker for early detection of age-related cognitive decline or conditions like Alzheimer’s
The image below shows a K-means clustering analysis of Age versus Clinical Dementia Rating (CDR) with K=2K=2K=2 clusters. This analysis is useful for observing patterns in cognitive impairment across different age groups. In this scatter plot, two clusters are represented: Cluster 0 (teal circles) and Cluster 1 (orange crosses). These clusters are defined by varying levels of CDR, which is a measure used to assess the severity of dementia.
Cluster 1 consists of individuals with a CDR of 0, indicating no dementia symptoms. These data points are distributed across a wide age range, spanning from approximately 40 to 90 years. This distribution suggests that individuals in this cluster are cognitively intact, regardless of age. The presence of these cognitively healthy individuals across such a broad age range highlights the variability in cognitive aging and suggests that age alone is not a definitive predictor of cognitive impairment.
Cluster 0 primarily includes individuals with higher CDR scores, ranging from 0.5 to 2.0. This cluster demonstrates increasing CDR scores with age, especially beyond 60 years. Higher CDR scores in Cluster 0 indicate a progression in cognitive impairment, with more individuals in older age groups showing higher levels of dementia. Notably, the presence of individuals with a CDR of 1.0 or 2.0 in older age brackets aligns with the understanding that dementia risk increases with age.
This clustering analysis shows a clear separation in cognitive states between the two clusters, with Cluster 1 primarily representing cognitively healthy individuals and Cluster 0 including those with varying levels of cognitive impairment. This pattern reinforces the notion that age is associated with increased dementia risk, but it also underscores the fact that dementia onset and progression are not uniform across all older adults. Some individuals maintain a CDR of 0 well into advanced age, while others show signs of cognitive decline starting in their 60s or 70s.
In summary, this clustering of Age and CDR provides insight into the heterogeneity of cognitive aging. By distinguishing between individuals with and without cognitive impairment across different age groups, the clustering offers a potential framework for identifying populations at higher risk of dementia, which could be beneficial for targeted early interventions and monitoring in clinical settings.
The image below displays a K-means clustering analysis of Age versus ASF (Atlas Scaling Factor) with K=2K=2K=2 clusters. This analysis aims to examine patterns in ASF values, a brain volume normalization factor, across different age groups. In this plot, Cluster 0 (represented by teal circles) and Cluster 1 (represented by orange crosses) indicate two distinct groupings in terms of ASF and age.
Cluster 0 generally includes individuals with lower ASF values, mostly below 1.3, and spans a broad age range from approximately 40 to 90 years. This cluster suggests that individuals with relatively lower ASF values, which may correspond to smaller normalized brain sizes, are distributed across various ages, without a strong age-specific pattern. This could indicate that lower ASF values are common across a range of ages, potentially representing a baseline or more typical brain structure among this population.
Cluster 1 predominantly contains individuals with higher ASF values, often exceeding 1.3, and is more concentrated in the age range of approximately 65 to 85 years. The pattern within Cluster 1 suggests that older individuals in this sample are more likely to have higher ASF values, which might correlate with age-related changes in brain volume. Since ASF adjusts for head size and potentially brain size, the higher ASF in older age groups could reflect structural changes that occur with aging.
The clustering is more influenced by ASF values than by age, as the clusters exhibit clearer separation along the ASF axis. This separation implies that ASF might be a significant factor in differentiating these groups, with age playing a secondary role in this distinction. Higher ASF values in Cluster 1 may suggest variations in brain structure that align with older age, which could have implications for understanding age-related brain changes or even risk factors related to cognitive decline.
In summary, this clustering of Age and ASF provides insights into how normalized brain size metrics vary across age groups. The two clusters highlight distinct ASF patterns that could be useful for identifying structural brain characteristics associated with aging. Further analysis could focus on how these clusters relate to cognitive measures, such as MMSE scores, to determine if higher ASF values in older individuals are linked to specific cognitive outcomes or risks. This clustering pattern suggests potential pathways for investigating the role of ASF in cognitive health assessments, particularly in the context of aging and neurodegenerative conditions like Alzheimer’s.
To further build on these findings, future steps could include exploring additional clusters (e.g., K=3 or K=4) to uncover more nuanced cognitive health subgroups. Additionally, investigating the correlation between ASF and other neurological measures could help clarify ASF’s role in cognitive health assessments. Overall, the clustering analysis sheds light on how cognitive health markers align with brain structure, offering potential pathways for early Alzheimer’s detection and personalized intervention strategies.
Polynomial Regression
The polynomial regression analysis between Estimated Total Intracranial Volume (eTIV) and Atlas Scaling Factor (ASF) demonstrates a strong, nearly linear relationship, with R2R^2R2 values consistently above 0.95 across polynomial degrees. While higher-degree polynomials (up to degree 4) show incremental improvements in R2R^2R2, the degree 1 (linear) and degree 2 (quadratic) models capture the relationship effectively, suggesting that increased complexity beyond a quadratic model offers little additional value and may even risk overfitting.
This tight correlation between eTIV and ASF suggests that ASF can be reliably derived from eTIV, a crucial insight for standardizing brain measurements across diverse populations. The connection between eTIV and ASF also supports using these metrics in models aimed at predicting cognitive health markers like MMSE (Mini-Mental State Examination) scores, given that brain structure and volume are linked to cognitive function. Since MMSE is a widely used measure for assessing cognitive impairment and early Alzheimer's, understanding how ASF and eTIV relate to MMSE could provide valuable predictive insights.
By confirming that ASF and eTIV scale predictably with brain size, we can leverage these features in machine learning models to predict MMSE scores more accurately. In this way, the relationship between eTIV, ASF, and MMSE may offer a standardized approach to assessing cognitive decline and could enhance the accuracy of early diagnosis models for conditions such as Alzheimer's disease.
Neural Network Performance
The results of the neural network training indicate that the model’s performance remained fairly stable over the 10 training epochs. Initially, the model’s training loss (Mean Squared Error, MSE) started at approximately 22.89, with a Mean Absolute Error (MAE) of 3.92. As training progressed, both the validation loss and MAE converged around values of 11.95 and 3.03, respectively. This stability in performance across epochs suggests that the model is learning some meaningful patterns but is not achieving a high degree of accuracy or substantial improvement with each epoch.
The close alignment between validation and training metrics suggests that the model is not overfitting, which is positive. However, the relatively high MAE of around 3 indicates that the model may not be effectively capturing the complexity of the patterns needed to predict MMSE scores with high accuracy. This could imply that the selected features (eTIV, nWBV, and ASF) may have limited predictive power for this specific task, or that further tuning of the model is necessary. Given the limited sample size of my data, this result is quite strong. With a larger dataset, the model’s performance could likely improve even further.
The chosen model architecture is relatively simple, consisting of two layers: a dense layer with 16 neurons and a ReLU activation function, followed by an output layer with a single neuron for regression. This simplicity helps prevent overfitting and enables faster convergence, as there are fewer weights to train. The model uses the Adam optimizer with MSE as the loss function and MAE as an evaluation metric, balancing efficient learning with a focus on minimizing absolute error. This setup provides a good foundation but may benefit from further tuning or additional layers with a larger data set to capture more complex patterns in the data.
To improve model performance, having a larger and cleaner dataset would be beneficial. The current dataset included a significant number of NaN values, which required extensive cleaning to ensure data quality and model accuracy. Addressing these missing values helped reduce noise, but a more comprehensive dataset could provide better feature representation, improve generalizability, and enhance predictive power.
Discussion:
This project applied a range of machine learning techniques—namely, clustering, polynomial regression, and neural networks—to detect Alzheimer’s risk using MRI biomarkers. Through K-means clustering, polynomial regression, and neural network analysis, distinct cognitive health profiles were identified, with the neural network showing particular promise in predicting cognitive decline. Each method contributed unique insights: K-means clustering helped segment patients based on cognitive profiles, polynomial regression underscored the value of simplicity in modeling relationships between biomarkers, and the neural network demonstrated high accuracy in handling complex data patterns relevant to clinical settings.
Key Takeaways:
K-means clustering provided a valuable framework for categorizing individuals into cognitive health profiles by clustering based on key brain and cognitive metrics. Through this analysis, it was evident that Mini-Mental State Examination (MMSE) scores were meaningfully related to both normalized Whole Brain Volume (nWBV) and the Atlas Scaling Factor (ASF). For instance, individuals with higher MMSE scores generally clustered together, often regardless of their ASF values, which suggests that ASF could complement MMSE as a metric for identifying diverse cognitive states. Additionally, clusters based on nWBV and MMSE illustrated a positive trend between brain volume and cognitive function, highlighting nWBV as a valuable metric in cognitive assessments.
Polynomial regression between Estimated Total Intracranial Volume (eTIV) and ASF further demonstrated that a simple linear model effectively captured their relationship, with high R² values even with minimal complexity. This outcome reinforces the utility of straightforward models in initial analyses to prevent overfitting and enhance interpretability.
Finally, the neural network model showed strong predictive performance for MMSE scores, which indicates its suitability for handling the nonlinear complexities in neuroimaging data. The model’s architecture and parameters achieved stable training and validation losses over multiple epochs, confirming that it could learn meaningful patterns without overfitting, making it a promising approach for clinical applications in Alzheimer's detection.
These insights collectively support the potential of machine learning in segmenting cognitive health profiles, guiding early diagnosis, and tailoring interventions for neurodegenerative diseases.
Limitations:
Some limitations impacted the results. The dataset size constrained the neural network’s ability to generalize to new data, affecting robustness and reliability. A larger, more diverse dataset would likely improve accuracy and allow the model to capture a broader range of cognitive health variations. Additionally, data imbalance, particularly in the distribution of lower MMSE scores, may have influenced sensitivity in detecting early dementia. A more balanced dataset or re-sampling techniques could enhance model sensitivity. While the dataset included derived biomarkers like eTIV, nWBV, and ASF, these may lack the detail found in raw MRI data, potentially limiting accuracy. The use of raw MRI images might provide richer details, which could enhance predictive accuracy, particularly for subtle patterns associated with early Alzheimer’s. Moreover, higher-degree polynomial models and neural networks showed a tendency toward overfitting, especially when capturing nonlinear relationships. Although regularization techniques mitigated this risk, simpler models or additional data could further enhance model generalizability.
Further Research:
Future research could build upon these findings through several approaches. Expanding the dataset to include a larger number of subjects with a broader range of cognitive scores would improve model robustness, particularly for neural networks, which benefit from extensive training sets. Analyzing raw MRI data could reveal subtler structural patterns associated with Alzheimer’s that derived biomarkers alone may miss, potentially enhancing the neural network’s ability to detect early cognitive decline. Developing additional features or exploring interactions among biomarkers and cognitive assessments could refine model accuracy, uncovering new predictive relationships and providing deeper insights into cognitive health markers. Exploring different neural network architectures, such as convolutional networks or ensemble learning approaches, could further enhance predictive accuracy without sacrificing interpretability.
Conclusion
This project demonstrates the potential of machine learning techniques, particularly neural networks, in early Alzheimer’s detection through MRI biomarkers and cognitive assessments. The clustering and polynomial regression analyses identified meaningful cognitive health markers, while the neural network achieved the highest predictive accuracy among the methods used. With larger, more diverse datasets and potentially raw MRI data, future research could further validate these findings and advance the development of machine learning tools for clinical applications in early Alzheimer’s detection.
Bibliography
Liu, V. X., Bates, D. W., Wiens, J., & Shah, N. H. (2019). The number needed to benefit: Estimating the value of predictive analytics in healthcare. Journal of the American Medical Informatics Association, 26(12), 1655–1659.
Zhang, Z. (2020). Predictive analytics in the era of big data: Opportunities and challenges. Annals of Translational Medicine, 8(4), 68.
Abudiyab, N. A., & Alanazi, A. T. (2022). Visualization Techniques in Healthcare Applications: A Narrative Review. Cureus, 14(11), e31355.
Ahsan, M. M., Luna, S. A., & Siddique, Z. (2022). Machine-learning-based disease diagnosis: A comprehensive review. Healthcare (Basel, Switzerland), 10(3), 541.
Lai, Y., et al. (2022). Identification of immune microenvironment subtypes for Alzheimer’s disease diagnosis and risk prediction based on explainable machine learning. Frontiers in Immunology, 13, 1046410.
Chen, Y.-C., et al. (2022). Predicting 30-day readmission for stroke using machine learning algorithms: A prospective cohort study. Frontiers in Neurology, 13, 875491. Assessed and Endorsed by the MedReport Medical Review Board